block-level metering may capture some positive externalities, but never captures any negative externalities.
This is not true.
The efficiency of the network is the efficiency with which transaction fees incentivize provision of whatever heterogenous computing services and data-flows are required to deliver the value that induces users to pay fees. Different nodes face different costs structures based on their relationship with the user and the block producer. Some of this is reflected in the technical processing differences you discuss.
Your underlying economic problem is that you don’t have efficient distribution of the fee based on the costs that different nodes could theoretically make to attract fees most efficiently, so nodes that can do valuable work can only protect themselves by restricting distribution in ways that maximize their income (“hoarding” and “private mempools”). Block-level payouts exacerbate this problem.
Transaction-level payments allow fees to reflect the granular and different cost-structures facing different users with heterogenous use cases. You would need to be able to modify the work function so that payouts are proportional to “work-contributed to fee collection” and this is not possible in the designs you are studying, which is why you see no downside.